Database Reference
In-Depth Information
These steps are not strictly necessary, but both steps are done in many cases for efficiency
purposes. Using RGB color images instead of grayscale increases the amount of data to be
processed by a factor of 3. Similarly, larger images increase the processing and storage
overhead significantly. Our raw 250 x 250 images represent 187,500 data points per image
using three color components. For a set of 1055 images, this is 197,812,500 data points.
Even if stored as integer values, each value stored takes 4 bytes of memory, so just 1055
images represent around 800 MB of memory! As you can see, image-processing tasks can
quickly become extremely memory intensive.
If we convert to grayscale and resize the images to, say, 50 x 50 pixels, we only require
2500 data points per image. For our 1055 images, this equates to 10 MB of memory,
which is far more manageable for illustrative purposes.
Tip
Another reason to resize is that MLlib's PCA model works best on
tall and skinny
matrices with less than 10,000 columns. We will have 2500 columns (that is, each pixel
becomes an entry in our feature vector), so we will come in well below this restriction.
Let's define our processing function. We will do the grayscale conversion and resizing in
one step, using the
java.awt.image
package:
def processImage(image: BufferedImage, width: Int, height:
Int): BufferedImage = {
val bwImage = new BufferedImage(width, height,
BufferedImage.TYPE_BYTE_GRAY)
val g = bwImage.getGraphics()
g.drawImage(image, 0, 0, width, height, null)
g.dispose()
bwImage
}
The first line of the function creates a new image of the desired width and height and spe-
cifies a grayscale color model. The third line draws the original image onto this newly
created image. The
drawImage
method takes care of the color conversion and resizing
for us! Finally, we return the new, processed image.
Let's test this out on our sample image. We will convert it to grayscale and resize it to 100
x 100 pixels:
